swe-bench

built from SWE-bench

Repository-level bug fixing: read the failing test, navigate an unfamiliar codebase, and produce a patch that resolves the issue.

Evidence

What this model claims, and what it has not measured yet.

accuracy vs frontier+1.2% (92.9% vs 91.7%)

cost saving vs frontier56% ($0.98 vs $2.22 per run)

latency saving vs frontier51% (p50 per run)

benchmarkSWE-bench

codepatchingrepositories

Measured on real benchmark episodes against the Claude Fable 5 frontier anchor. Accuracy is the task-success gap in points; savings are per run, latency as p50 model time.

Pipeline

How far this model has come.

Traces capturedNo trace corpus is published for this model yet.
Simulation builtNo simulation is published for this model yet.
Optimization evaluatedCandidate models scored against the scenarios; the winner serves.

Run it

Everything that changes something, or spends anything, is behind the login.

Sign in to open the playgroundSend this model real requests and watch the responses come back.Sign in to serve this modelGet an OpenAI-compatible URL and an API key for your own workload.

Models